On Fri, 2006-04-07 at 19:05 -0400, Tom Lane wrote:
> It's plausible though that we are seeing contention across members of
> the LWLock array, with the semop storm just being a higher-level symptom
> of the real hardware-level problem. You might try increasing
> LWLOCK_PADDED_SIZE to 64 or even 128, see
> src/backend/storage/lmgr/lwlock.c (this is something that does exist in
> 8.1, so it'd be easy to try).
pSeries cache lines are 128 bytes wide, so I'd go straight to 128.
If you're renting all 8 CPUs, I'd drop to 4 and try that instead. With 8
CPUs the contention will vary according to what each CPU is doing at any
one time - when they all hit the contention spot, things will get worse.
The pSeries has good CPUs and great caching, so I'd expect contention to
be somewhat more apparent as a bottleneck.
Best Regards, Simon Riggs
